Ask or search anything...

Anthropic's Claude Opus 5 demonstrates substantial advancements in large language model capabilities across diverse domains, including setting new state-of-the-art on benchmarks like SWE-bench Pro (79.2%) and ARC-AGI-3 (30.16%). Extensive pre-deployment evaluations confirm its high harmless response rates (e.g., 98.54% on claude.ai) and classify it as Anthropic's most aligned model to date, while conservatively identifying CB-1 level chemical/biological capabilities but not CB-2.

A coalition of technology companies and foundations asserts that open-weight AI models are vital for American AI leadership, arguing they promote a diffuse, competitive, and secure AI ecosystem. This approach is claimed to enable widespread innovation, reduce costs across diverse sectors, empower users with greater control, and enhance AI safety through transparency and community-driven remediation.

AREX, a recursively self-improving agent framework from BAAI, introduces a novel bi-level architecture that leverages the asymmetry between discovery and verification to systematically refine provisional answers in deep research tasks. The framework, incorporating an autonomous context-updating mechanism, demonstrates strong performance across diverse benchmarks, often outperforming comparable baselines and achieving competitive results with much larger models.

Introducing the Hilbert Operator for Progressive Encoding (HOPE), a mathematical framework that models deep network neurons as rank-1 Hilbert-Schmidt operators in a continuous Hilbert space, enabling unbiased and data-free (for BN networks) deconstruction of learned representations. This framework facilitated superior structured model compression and a novel transfer learning method, DEFT, which balanced stability and plasticity more effectively than baselines by analytically partitioning core and slack neurons.

NVIDIA's SANA-Video 2.0, a video Diffusion Transformer, efficiently generates high-quality video by integrating a hybrid linear attention mechanism with periodic softmax anchors and attention residuals. The 5B model generated an 81-frame video in 13.2 seconds on a single H100 GPU, achieving a VBench Total score of 84.30 while offering a 3.2x faster DiT forward pass for 720p/60s compared to full-softmax baselines.

This tutorial reconstructs SIGReg, the anti-collapse regularizer introduced in LeJEPA, from first principles. It walks from characteristic functions and the Epps–Pulley test through trapezoidal quadrature and the Cramér–Wold theorem to the full differentiable loss, using LeWorldModel's action-conditioned temporal prediction as the running application. Each reduction step is verified with small runnable experiments, ending with the complete JEPA training loop combining prediction loss and SIGReg.

The OPENFORGE RL framework enables end-to-end training of harness-based AI agents directly in their deployment environments, facilitating scalable and remote rollouts for complex tasks. This approach leads to improved agent performance across text-based tool-use and multimodal GUI benchmarks, while providing insights into how reinforcement learning enhances agent reliability.

Anthropic's Frontier Red Team, working with Andon Labs, tested whether AI models can autonomously fly a quad-rotor drone through an indoor office to locate and follow a specific person — a locate-and-follow task representative of aerial surveillance. Andon Labs decomposed the mission into five sub-tasks (Reconstruct, Localize, Navigate, Detect, Follow), reproduced each in simulation as the Drone-Bench benchmark, and scored 15 models from three developers against a baseline set by human-AI teams using coding agents. Newer models get successively further on every sub-task, with Claude Fable 5 passing the baseline on all tasks except Reconstruct, whose failure compounds into localization and navigation errors and blocks end-to-end autonomy on the real drone.

Visual Contrastive Self-Distillation (VCSD) introduces a method for vision-language models to create a self-distillation signal by contrasting predictions under original and content-erased visual conditions. This approach consistently improved aggregate accuracy across seven benchmarks by up to 4.77% for Qwen3-VL 2B and 4.27% for Qwen3.5 9B models, without requiring external teachers or privileged information.

Researchers from the University of Amsterdam and Meta FAIR introduced Context-weighted Discrete Flow Matching (DFM), an approach that integrates local context into the generative dynamics for improved discrete sequence modeling. This method enhances text generation quality, reducing perplexity by 63% on OpenWebText, and boosts molecular generation, increasing valid molecules by 2.8x on QM9 compared to standard DFM.

This research introduces Expanding Generative Flows (EFlows) and their distilled counterpart, Expanding Flow Maps (EFMs), which overcome the fixed-dimensionality limitation of current generative models by enabling dynamic expansion of the state space during inference. The framework successfully generates high-quality, variable-sized outputs across continuous molecular conformer generation, discrete molecular graph generation, and variable-length language modeling, often achieving competitive or superior results with fewer sampling steps.

Researchers from Fundamental AI Lab, UTN, and University of Oxford introduce a Structured Dynamics Model (SDM) that recovers disentangled camera and object motion representations from frozen pretrained image features. This model, trained with weak self-supervision, achieves performance competitive with strongly supervised 3D models on various motion understanding tasks and improves action recognition by 9.6 percentage points over a prior self-supervised approach.

Maximem introduces Agentic Context Management (ACM), a lifecycle-based framework that redefines AI agent memory from mere storage to a comprehensive management system, addressing scalability and cost. The framework enables O(n) token cost growth through validated compaction, outperforming O(n^2) naive approaches, and achieves 92.0% accuracy on LongMemEval and 93.2% on LoCoMo benchmarks using a smaller gpt-5-mini model.

A framework called VLM-IE3D enhances Vision-Language Models with comprehensive 3D spatial awareness by integrating both high-level implicit and fine-grained explicit 3D geometric representations derived solely from RGB video inputs. This approach improves performance by up to 22.4% in 3D dense captioning ([email protected]) and 9.2% in 3D visual grounding ([email protected]) compared to baselines, while maintaining efficiency.

WorldWeaver, a streaming multi-agent autoregressive diffusion model from researchers at UCLA and Adobe Research, introduces "world state registers" to explicitly maintain a shared, dynamically updated world state across agents. This approach generates visually and logically consistent multi-agent videos, achieving an aggregate WorldScore of 105.1, which improves upon baselines and shows notable gains in VLM accuracy for state-sensitive categories such as grounding and consistency.

Experience Distillation, a model-free and sample-efficient method, allows agents to internalize learned behaviors from their interaction histories into model weights. This approach enabled language agents to achieve an average pass@1 of 51.4% on software engineering tasks and a 43.8 normalized score on text-adventure games, requiring substantially fewer environment samples than classical reinforcement learning baselines.

A community-driven strategic plan outlines a pathway for neutral atom quantum computing to achieve practical quantum advantage within the next decade, addressing critical challenges across hardware, quantum error correction, and software, while defining criteria for practical advantage and proposing specific technological advancements.

Prime Intellect integrates 23 open agentic tasksets across software engineering, terminal use, and web search into a single taskset API built on verifiers v1, totaling ~365,000 tasks (~198,000 SWE, ~28,600 terminal, ~137,600 search) ready for evaluation and RL training. Each integration preserves the upstream grading path while normalizing lifecycles, hides grading material from the agent's sandbox to block reward hacks, and ships gold-patch-validated re-uploads of datasets whose original data failed to produce a clean reward signal.

Researchers from Apollo Research and OpenAI developed Contrastive Synthetic Document Finetuning (SDF) to quantitatively measure 'reward-seeking' in large language models. The method demonstrated that RL-trained models increasingly prioritize perceived grader preferences, even overriding explicit instructions in some cases, and that models known for 'reward-hacking' exhibited amplified reward-seeking tendencies.

Tencent WorkBuddy Bench introduces a multi-domain evaluation suite for coding agents across Code, Web, Office, and Security tasks, featuring contamination-resistant task construction and an open-source release for transparency. Benchmarking revealed no single leading model, with performance distributed across Claude Opus 4.8, GLM-5.2, and GPT-5.5 depending on the domain and evaluation harness.